Tencent releases WeVisOne-VL for detection, segmentation and alpha matting

The Apache-2.0 vision-language model combines object- and pixel-level perception in downloadable 2B and 4B checkpoints, with publisher-reported gains across detection, OCR and matting.

2 min read

A computer monitor showing a photo-editing interface, illustrating AI-assisted image generation and editing.

A computer monitor showing a photo-editing interface, illustrating AI-assisted image generation and editing.

Tencent’s WeChat Vision team has released WeVisOne-VL, an Apache-2.0 vision-language model designed to handle object-level and pixel-level perception in one system. The release includes downloadable 2B and 4B checkpoints, an inference package, tutorials, a demo and public evaluation code.

Built on Qwen3-VL, WeVisOne-VL adds a 9.6-million-parameter pixel decoder. It supports open-vocabulary detection, referring localization, counting, pointing and OCR, as well as referring segmentation and promptable alpha matting. That lets the model return points, boxes, binary masks or continuous alpha mattes depending on the task.

Tencent reports that the 4B model scored 71.3 F1 on Dense200 detection, compared with 58.7 for the strongest listed baseline; 52.0 F1 on TotalText OCR, versus 43.3; and 71.2 gIoU on ReasonSeg, versus 65.0. On the team’s WeRefMat referring-matting benchmark, it recorded 9.8 mean absolute deviation, where lower is better, compared with 22.4 for the best listed baseline. These are publisher-reported results using greedy decoding and no more than 1,960 visual tokens per image, not independent replications.

The model card describes the release as a single-turn image-perception system. Its pixel decoder does not accept video grids, and fine boundaries can remain sensitive to resolution, occlusion, transparency and prompt phrasing. No hosted Hugging Face inference provider is currently listed, so users need local tooling such as Transformers, vLLM or SGLang.

The canonical Tencent repository showed 83 total stars, three forks and two commits when checked on October 24, 2026. That is a popularity snapshot, not measured growth; no reliable historical star baseline was available.

Image: JJ ROCHA / Unsplash.

vision-language modelsTencentWeVisOne-VLcomputer visionopen models